Johns Hopkins University
The Beer Lab
Regulatory Genomics
Dr. Michael Beer is a Professor in the Department of Biomedical Engineering at Johns Hopkins University, with secondary appointments in Genetic Medicine, Oncology, and Molecular Biology and Genetics. He earned his PhD and MA in Astrophysical Sciences from Princeton University and a BSE from the University of Michigan.
The Genome's Operating System
Every cell in the human body runs on identical DNA. At Johns Hopkins, Professor Michael Beer and his lab are using machine learning, functional genomics, and CRISPR-based experiments to decode the regulatory sequences that decide which genes get switched on, and why simply reading the genome was never going to be enough.
By BioBuilt Editorial
The human genome contains roughly 3.2 billion letters of DNA. Only about one to two percent of them code for proteins, the molecules that build and run a cell. For decades, that left biologists with an uncomfortable question: what is the other 98 percent of the genome actually doing?
The early answer, popular for much of the twentieth century, was "not much." Non-coding DNA was widely dismissed as junk, evolutionary leftovers with no real function. That view hasn't survived contact with modern genomics. Buried inside that supposedly empty majority of the genome are millions of regulatory sequences called enhancers, short stretches of DNA that don't build anything themselves but instead decide when a nearby gene turns on, in which cell type, and how strongly. If genes are the genome's parts list, enhancers are closer to its operating system, the software running underneath, deciding which parts actually get used and when.
At Johns Hopkins, biomedical engineer and geneticist Michael Beer has spent nearly two decades trying to decode that operating system, work that sits at the intersection of computer science, physics, and molecular biology.
The Parts Are Easy to Find. The Instructions Aren't.
Finding a gene inside a strand of DNA is, by the standards of modern genomics, a largely solved problem. Genes follow recognizable rules: long, uninterrupted stretches of three-letter codons, clearly defined start and stop signals, and well-characterized splice sites where the cell edits its own genetic message before using it. Decades of RNA-sequencing experiments have also produced enormous datasets confirming exactly where real genes begin and end, giving computational biologists both a rulebook and an answer key.
"Protein coding genes have known properties that aid the computational detection of genes," Beer explained, pointing to those consistent codon structures and splice signals. "On the other hand, we do not yet know the sequence properties of regulatory enhancers at that level of detail."
That gap is not for lack of trying. The Encyclopedia of DNA Elements project, known as ENCODE, launched in the early 2000s specifically to catalog the genome's functional parts beyond its genes, and Beer's lab, backed by a $1.8 million NIH grant, has contributed to that effort directly. But enhancers resist the kind of clean rulebook that exists for genes. They're typically only a few hundred base pairs long, and they sit scattered across enormous stretches of DNA that do nothing detectable at all, more like single sentences hidden somewhere inside an otherwise blank book than a clearly labeled chapter. Worse, an enhancer's function isn't written in one obvious signal. It emerges from combinations, specific clusters of short sequences called transcription factor binding sites, which only work together, in particular arrangements, to switch a gene on. Change the combination, or the spacing between the pieces, and the same DNA can go from active to silent.
Experimental tools capable of directly measuring which stretches of DNA are functioning as enhancers in a given cell type, technologies like ATAC-seq and DNase-seq that reveal which regions of chromatin are open and accessible, have only existed for a little over a decade. Beer's lab has spent roughly twice that long building computational methods, including kmer-based support vector machines and deep neural networks, that scan raw DNA sequence for the short, recurring combinations of letters statistically associated with enhancer activity. It's not unlike training a spam filter to recognize which combinations of words tend to show up in junk email, except the "words" are four-letter DNA fragments, and the cost of a wrong answer isn't a missed advertisement, it's a mischaracterized region of the genome.
"We have made tremendous progress over the last 20 years," Beer said, "but much is still unknown."
Why It Matters
The reason any of this is worth two decades of a scientist's career becomes clear once you sit with a strange fact: every cell type in the human body is running the exact same genome. A neuron and an immune cell don't differ because they carry different DNA. They differ because they've switched on almost entirely different sets of genes.
"Each of the different cell types in the human body has the same DNA but different cell types perform completely different, and often essential, functions," Beer said. "The process of the development of a human from a simple embryo is completely controlled by DNA regulatory sequences."
That control isn't a side detail of development, it is development: the whole process of a single fertilized cell dividing and specializing into a heart, a brain, and everything else depends on regulatory DNA switching the right genes on in the right cells at the right moments, and keeping the wrong ones off. It's also, unsurprisingly, where a great deal of disease originates. Genome-wide association studies have repeatedly found that most disease-linked genetic variants in humans sit outside protein-coding genes entirely, inside these regulatory regions. A single mutation in an enhancer, sometimes located hundreds of thousands of DNA letters away from the gene it controls, can be enough to misfire a developmental signal or push a healthy cell toward cancer, without ever touching a gene directly.
An Engineer's Approach to Biology
Beer didn't arrive at this problem through a conventional biology track. His PhD is in astrophysical sciences from Princeton, and his doctoral and postdoctoral research modeled the physics of controlled nuclear fusion, work recognized with the Simon Ramo Award and a Department of Energy postdoctoral fellowship. When asked what carried over from modeling plasma physics into modeling gene regulation, he didn't offer a quick answer. He pointed instead to an entire paper he had written on the subject, a 2021 essay in the journal Current Genomics addressed to other physicists and engineers weighing the same career leap, in which he describes his physics training as excellent preparation for nearly every part of the transition.
That instinct, treat a complicated system as something to be modeled and tested rather than simply cataloged, shows up directly in how his lab operates. Rather than compiling lists of where transcription factors bind, Beer's team builds predictive models trained to learn the actual combinatorial logic: which specific groupings of binding sites, in which arrangement, make an enhancer active in a liver cell but silent in a neuron. It's the difference between memorizing a phrasebook and reverse-engineering the grammar underneath it.
Testing Predictions Against Biology
A model that fits existing data well isn't automatically a model that has learned something true, and Beer is exacting about that distinction. His lab pairs its computational predictions with direct experimental tests, most notably CRISPR interference, a technique that uses a modified, "catalytically dead" version of the CRISPR-Cas9 system to silence a suspected enhancer without cutting the DNA at all, then measures what happens to nearby gene activity as a result. If a model predicts a sequence is a functioning enhancer, CRISPR interference is how you find out whether reality agrees.
In 2023, Beer's lab published a study in Nature Genetics using exactly this approach, a CRISPR interference screen that traced how enhancers are physically brought into contact with the genes they regulate through loops in the genome's structure, anchored by a protein called CTCF, and how those loops constrain and shape enhancer activity as a cell transitions from one state to another. It's a finding that only becomes possible once a computational model has narrowed down which sequences are worth testing in the first place, and only becomes trustworthy once that prediction has survived a real experimental check.
"Computational predictions are trained on one set of data," Beer said. "Testing on a separate set of experiments evaluates how well the computational model extrapolates or generalizes to other data. This is absolutely crucial, since modern computational models have millions of free parameters and can basically fit anything."
That last point is worth sitting with. A sufficiently large model can often fit almost any dataset handed to it, whether or not it has learned anything real about the underlying biology. The only way to tell the difference is to make the model commit to a prediction about data it has never seen, then go back into the lab and check.
His group's current cancer research runs on a version of that same cycle: design an experiment in cancer cells to generate relevant data, train a model on it, use the model's predictions to identify candidate regulatory sequences, genes, or drug responses, then return to the lab to test those specific predictions, sometimes against new drugs entirely, and use what's learned to design the next round of experiments.
"This is just how science works in general," Beer said.
A Skeptic of His Own Field's Hype
Because Beer helped build the machine learning toolkit now standard across regulatory genomics, his read on the field's current AI enthusiasm carries some weight, and it is noticeably more cautious than the surrounding hype cycle.
"I think machine learning will be absolutely necessary to build a complete theory of regulatory sequence function," he said. "But I also think that just hoping that you can throw a bunch of data into a large neural network and get out anything close to a correct understanding of gene regulation is wishful and dangerous thinking."
It's a pointed distinction. Scale alone, in his view, isn't a substitute for the harder discipline of building models with real, testable structure, and then subjecting them to exactly the kind of experimental scrutiny his lab applies through CRISPR interference. A neural network that can't be checked against biology isn't understanding, however good its predictions look on paper.
The Road Ahead
Beer isn't promising a fast resolution. By his own estimate, it will take another ten to twenty years before scientists have a genuinely clear map of where enhancers sit in the genome and how they function. But the direction of that map is already visible in his lab's recent work. A 2023 study from his group built a machine learning model capable of predicting which enhancers control the transitions between cell types, a step that researchers hope could eventually enable therapies that switch specific genes on or off with real precision, rather than the blunter, system-wide effects of most conventional drugs.
Beer is candid about how large he thinks the eventual payoff could be.
"It is probably not an exaggeration to say that if we completely understand how the regulatory sequences controlling human development and disease function," he said, "we will be able to cure cancer and design new organisms."
The Human Genome Project answered the question of what letters make up human DNA. The harder, still-unfinished question, the one Beer has spent his career on, is what those letters are actually being told to do, and by whom.